refactor: improve prompt

app.py

@@ -57,30 +57,37 @@ def load_dataset(query):
     })
     return df
 
-def generate_answer(question, context, max_length=150):
+def generate_answer(question, context, max_length=300):  
     tokenizer, model = load_models()
     
-    # Improve prompt to generate concise, summarized answers
-    prompt = f"""Based on scientific research about autism, provide a brief, clear summary answering the following question. 
-    Focus only on the most important findings and be concise.
+    # Enhanced prompt for more detailed and structured answers
+    prompt = f"""Based on scientific research about autism, provide a comprehensive and structured summary answering the following question.
+    Include the following aspects when relevant:
+    1. Main findings and conclusions
+    2. Supporting evidence or research methods
+    3. Clinical implications or practical applications
+    4. Any limitations or areas needing further research
+    
+    Use clear headings and bullet points when appropriate to organize the information.
     If the context doesn't contain relevant information about autism, respond with 'I cannot find specific information about this topic in the autism research papers.'
     
     Question: {question}
     Context: {context}
     
-    Provide a concise summary:"""
+    Detailed summary:"""
     
     # Optimize input processing
-    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512)
+    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=768)  
     
     with torch.inference_mode():
         outputs = model.generate(
             **inputs,
             max_length=max_length,
-            num_beams=2,
-            temperature=0.7,
+            num_beams=4,  
+            temperature=0.8,  
             top_p=0.9,
-            repetition_penalty=1.2,
+            repetition_penalty=1.3,  
+            length_penalty=1.2,  
             early_stopping=True
         )
     
@@ -90,11 +97,13 @@ def generate_answer(question, context, max_length=150):
     if torch.cuda.is_available():
         torch.cuda.empty_cache()
     
-    # Additional validation of the answer
+    # Enhanced answer validation and formatting
     if not answer or answer.isspace() or "cannot find" in answer.lower():
         return "I cannot find specific information about this topic in the autism research papers."
     
-    return answer
+    # Format the answer with proper line breaks and structure
+    formatted_answer = answer.replace(". ", ".\n").replace("• ", "\n• ")
+    return formatted_answer
 
 # Streamlit App
 st.title("🧩 AMA Autism")